---

Why Say the Word

Voice typing is one of the biggest productivity tools on a computer, and for people with RSI, carpal tunnel, dyslexia, limited mobility or chronic pain it may be the only comfortable way to write. The best dictation tools usually send every word you say to someone else's server, on a subscription. The ones built into the OS are private, but give you little control.

Say the Word gives you fast hold-to-talk dictation in every app, with every step running on your own Mac:

• Private by design. No account, no cloud, no telemetry. Your voice and your text stay on your Mac.
• Fast. Verbatim mode took a median of about 0.4 seconds from key release to the paste request in synthetic-speech tests on an M4 Mac with 16 GB. See the measurements below.
• Works everywhere you type. Mail, Slack, Notes, your browser, your code editor.
• Works offline. After a one-time model download, it needs no internet.
• Open source. MIT-licensed: anyone can read the code and check every privacy claim.

How it compares

Cloud dictation apps such as Wispr Flow and Say the Word make different trade-offs. As of October 2026, from Wispr Flow's own pricing, privacy and language pages:

If you want many languages, your phone, or team features, Wispr Flow is the stronger choice. If your words must never leave your Mac, or you want to read the code that handles them, choose Say the Word.

Features

• Hold to talk. Hold Fn, speak, let go: the text is pasted at your cursor.
• Hands-free. Press Fn twice (or Fn+Space, or click the pill) and speak with no key held. Press Fn again to stop.
• Cleaned mode. Rules and an optional local Ollama model remove "um", false starts and repeated words. Checks look for unwanted edits, including changed numbers and negations. If a check flags the result or the model takes too long, the app falls back to rules-only cleanup. You can compare the original and cleaned text in History.
• Verbatim mode. The speech recognizer's original transcript, with punctuation and capitals.
• Recover missed pastes. Esc cancels and Undo brings the dictation back while it is offered. Cmd+Ctrl+V pastes your last available dictation again. If a paste can't happen, the pill offers Copy. Recovery stays in memory until you quit.
• Safe with passwords. When macOS reports a password field, the app never pastes automatically.
• Your choice of key. No Fn key, or another app already uses it? Use Control+Option instead.
• History you control. See what was heard beside what was written, and how long each step took. Kept in memory only, unless you choose to save it.
• Made for everyone. Light and dark appearance, a guided first run, and support for VoiceOver, Increase Contrast, Reduce Motion and Reduce Transparency. Everything also works from the menu bar.

Privacy, in plain terms

These rules are enforced in code and covered by tests: see Architecture and behaviour.

Install

Download Say the Word for macOS 14 or later on Apple Silicon (M1 or later). It is signed and notarized by Apple.

1. Open the DMG and drag Say the Word into Applications.
2. Open it from Applications. It lives in the menu bar, and a setup guide opens on the first launch.
3. Allow Microphone and Accessibility access when asked (Accessibility lets it paste for you). The speech model downloads once, about 670 MB.
4. Click into any text field, hold Fn, speak, and let go.

Optional, for Cleaned mode: install Ollama, run ollama pull qwen3.5:4b, and pick the model on the Cleanup page.

Updates are not automatic yet. To update, download the latest release and replace the app; your settings and any history you chose to save stay. If you use another dictation app that listens to Fn, quit it first: two apps on the same key will fight over it.

FAQ

Is it really free? Yes. It is MIT-licensed, with no account, no subscription and no paid tier.

Does it work on Intel Macs? No. It needs Apple Silicon (M1 or later) and macOS 14 or later.

Which languages does it understand? The speech model, NVIDIA's Parakeet v3, recognizes 25 European languages, including English, French, German, Spanish and Italian. The cleanup rules are tuned for English.

A dictation didn't arrive. What happened? The pill says why in a few words, and the History page explains it, with what you can do about it. If the text was recognized, use Copy on the pill, or press Cmd+Ctrl+V to paste your last available dictation. Recovery stays in memory until you quit; it cannot recover every failure. The log (menu-bar icon › Show Log) has the details, and never contains what you said.

It says "Secure Input is on: nothing was pasted". macOS reports that the app in front is taking a password, so automatic paste is refused. If your cursor isn't in a password field, macOS or that app has left Secure Input on. Copy the text from the pill and paste it yourself.

How it works

flowchart LR
    K["Fn key<br/>(Swift helper)"] --> M["Main process<br/>(TypeScript)"]
    M --> O["Overlay pill<br/>(microphone)"]
    O -- 16 kHz audio --> S["Speech worker<br/>Parakeet v3<br/>via sherpa-onnx"]
    S -- text --> C{"Cleaned?"}
    C -- no --> P
    C -- yes --> R["Rules + local Ollama<br/>+ guard"] --> P["Swift helper:<br/>paste into the<br/>app you were in"]

• Electron and TypeScript hold all the product logic. A small Swift helper does what JavaScript can't: seeing the Fn key, posting the paste, reading Accessibility.
• Speech recognition: Parakeet TDT 0.6b v3 (int8) on the CPU through sherpa-onnx, in its own process, with Silero voice activity detection.
• Cleanup: fast rules first, then an optional local model with a fixed time limit and a guard that checks for common unwanted changes. The guard is heuristic: it cannot guarantee that meaning is preserved, so the original transcript stays available for comparison.
• Storage: history in SQLite in a separate process, so a slow disk never delays a paste.

Measured on a MacBook Air with an Apple M4, 16 GB and macOS 26.6 (benchmarks). The dictation timings use synthetic speech and intercepted paste requests; they do not measure accuracy on real voices or the time taken by every destination app. Results vary with your Mac, recording and cleanup model.

Build from source

You need macOS 14+ on Apple Silicon, Node 22.13+, and the Xcode command-line tools (Swift 6).

git clone https://github.com/yravinderkumar33/say-the-word.git
cd say-the-word
npm install
npm run models:download   # the speech model, about 670 MB, once
npm run dev               # builds the Swift helper and starts the app with hot reload

In development, macOS gives the Microphone and Accessibility permissions to the terminal that started the app. To test them as a user would, build a signed app: npm run setup:signing once, then npm run pack, then open dist/mac-arm64/Say the Word Dev.app.

The design and the decisions behind it are in docs/: research and decisions, architecture, implementation reference, and the progress log.

Roadmap

• A "Check for updates" button that respects the network policy
• More languages through a Whisper-based engine
• Windows and Linux (the key, paste and speech parts are already behind interfaces)

Contributing

Contributions are welcome, especially:

• Testing on other Macs, keyboards and apps. Paste behaviour differs between editors, browsers and terminals. A log excerpt and the app's name help a lot.
• Accessibility feedback from people who use VoiceOver or rely on dictation every day.
• Better cleanup rules, and better ways to measure Cleaned mode.

Open an issue for a bug or an idea. Before a pull request, run npm run check and read the working agreements in CLAUDE.md. The privacy rules are requirements: never log what was said, use no network beyond the local Ollama and downloads the user starts, and copy no code from GPL or AGPL projects.

Acknowledgements

Say the Word builds on excellent open work: sherpa-onnx, NVIDIA's Parakeet, Ollama, Electron, React and Tailwind CSS. The native layer learned from these MIT-licensed dictation projects: Amical, OpenWhispr, Handy, hive and dictation-cleanup-rules. Every component and its licence is listed on the app's About page.

Say the Word is an independent project, not affiliated with Wispr Flow or any other dictation product.

Licence

MIT